def test_per_channel_qtensor_creation(self):
numel = 10
ch_axis = 0
scales = torch.rand(numel)
zero_points_int = torch.randint(0, 10, size=(numel, ))
zero_points_float = torch.randn(numel)
for dtype, zero_points in itertools.product(
[torch.qint8, torch.quint8], [zero_points_float, zero_points_int]):
q = torch._empty_per_channel_affine_quantized(
[numel],
scales=scales,
zero_points=zero_points,
axis=ch_axis,
dtype=dtype)
# TODO(#38095): Replace assertEqualIgnoreType. See issue #38095
self.assertEqualIgnoreType(scales, q.q_per_channel_scales())
self.assertEqual(zero_points, q.q_per_channel_zero_points())
self.assertEqual(ch_axis, q.q_per_channel_axis())
# create Tensor from uint8_t Tensor, scales and zero_points
for zero_points in [zero_points_float, zero_points_int]:
int_tensor = torch.randint(0,
100,
size=(numel, ),
dtype=torch.uint8)
q = torch._make_per_channel_quantized_tensor(
int_tensor, scales, zero_points, ch_axis)
self.assertEqual(int_tensor, q.int_repr())
# TODO(#38095): Replace assertEqualIgnoreType. See issue #38095
self.assertEqualIgnoreType(scales, q.q_per_channel_scales())
self.assertEqual(zero_points, q.q_per_channel_zero_points())
self.assertEqual(ch_axis, q.q_per_channel_axis())
def quantized_tensor_to_pytorch(tensor: torch.Tensor,
scale,
zp,
num_bits,
mode,
dest_dtype,
per_channel=False,
channel_dim=0):
"""
Convert a tensor quantized with quantization parameters calculated by CACP to a PyTorch "native" quantized
tensor.
We refer to quantization parameters calculated using either of:
* quantization.symmetric_linear_quantization_params
* quantization.asymmetric_linear_quantization_params
And to tensors quantized using either of:
* quantization.linear_quantize
* quantization.linear_quantize_clamp
Args:
tensor (torch.Tensor): The tensor quantized in CACP
scale (torch.Tensor): Scale factor calcualted by CACP
zp (torch.Tensor): Zero point calcualted by CACP
num_bits (int): Number of bits used for quantization in CACP
mode (quantization.LinearQuantMode): The quantization mode used in CACP
dest_dtype (torch.dtype): PyTorch quantized dtype to convert to. Must be one of: torch.quint8, torch.qint8
per_channel (bool): Flag in indicating if tensor was quantized per-channel
channel_dim (int): If per_channel is set, this indicates the dimension of the channel in the tensor
Returns:
PyTorch quantized tensor (dtype one of torch.quint8 / torch.qint8 / torch.qint32)
"""
assert (
tensor == tensor.int()).all(), 'Tensor does not appear to be quantized'
converted_scale, converted_zp = qparams_to_pytorch(scale,
zp,
num_bits,
mode,
dest_dtype,
reduce_range=False)
zp_diff = -converted_zp.view(zp.shape) - zp
if dest_dtype == torch.quint8:
temp_dtype = torch.uint8
elif dest_dtype == torch.qint8:
temp_dtype = torch.int8
else: # dest_dtype == torch.qint32:
temp_dtype = torch.int32
tensor = (tensor - zp_diff).to(temp_dtype)
if per_channel and scale.shape[channel_dim] > 1:
return torch._make_per_channel_quantized_tensor(
tensor, converted_scale, converted_zp, channel_dim)
return torch._make_per_tensor_quantized_tensor(tensor, converted_scale,
converted_zp)
def test_per_channel_qtensor_creation(self):
numel = 10
ch_axis = 0
scales = torch.rand(numel)
zero_points = torch.randint(0, 10, size=(numel,))
q = torch._empty_per_channel_affine_quantized(
[numel], scales=scales, zero_points=zero_points, axis=ch_axis, dtype=torch.quint8)
self.assertEqual(scales, q.q_per_channel_scales())
self.assertEqual(zero_points, q.q_per_channel_zero_points())
self.assertEqual(ch_axis, q.q_per_channel_axis())
# create Tensor from uint8_t Tensor, scales and zero_points
int_tensor = torch.randint(0, 100, size=(numel,), dtype=torch.uint8)
q = torch._make_per_channel_quantized_tensor(int_tensor, scales, zero_points, ch_axis)
self.assertEqual(int_tensor, q.int_repr())
self.assertEqual(scales, q.q_per_channel_scales())
self.assertEqual(zero_points, q.q_per_channel_zero_points())
self.assertEqual(ch_axis, q.q_per_channel_axis())
def _clamp_weights(qweight, observer, scale, zp):
if not _needs_weight_clamping(observer, qweight.dtype):
return qweight
observer = _get_weight_observer(observer)
min_, max_ = observer.quant_min, observer.quant_max
# Doing this because can't use torch.ops.quantized.clamp() with per_channel qscheme yet.
qw_int_max = torch.clone(qweight.int_repr()).fill_(max_)
qw_int_min = torch.clone(qweight.int_repr()).fill_(min_)
qw_int = torch.minimum(torch.maximum(qweight.int_repr(), qw_int_min), qw_int_max)
if observer.qscheme in [torch.per_tensor_symmetric,
torch.per_tensor_affine]:
qweight = torch._make_per_tensor_quantized_tensor(qw_int, scale.item(), zp.item())
elif observer.qscheme in [torch.per_channel_symmetric,
torch.per_channel_affine,
torch.per_channel_affine_float_qparams]:
qweight = torch._make_per_channel_quantized_tensor(qw_int, scale, zp, axis=observer.ch_axis)
else:
raise ValueError("Unexpected qscheme " + observer.qscheme)
return qweight
